AI systems
Embedded and on-device inference
Small models on microcontrollers and sensor boards that classify or detect without sending raw data.
We design, integrate and deploy.
MLPerf Tiny · ONNX specification
/
Frontier labs/
Into the productFrom a result to a qualified release.
Into the product, inside Frontier labs: embedded and on-device inference, edge server inference, quantisation, distillation, pruning and sparsity and more. Every item is traced to its specification, from MLPerf Tiny to ONNX specification, and runs on the open Observatory Protocol.
Plan with usArchitectureThe protocol- Technologies
- 9
- At the frontier
- 1
- Protocol functions
- 3 of 8
- Industries served
- 3
Reference design, drawn to scale · drag to turn · hover a part
01What it is
Edge inference, model serving and optimisation.
AI from model to accountable operation: edge inference and model optimisation, foundation models for Earth observation and geospatial work, time-series and sensor AI, multimodal fusion and supervised agentic operations, federated and privacy-preserving learning, physical AI and robotics within stated red lines, MLOps, evaluation and model risk, and the governance frameworks that apply to each. Each technology below carries what it is, what we do with it, and the standard it is drawn against.
AI systems
Edge server inference
Vision, language and sensor models served from far and near edge servers at bounded latency.
We design, integrate, deploy and operate.
MLPerf Inference: Edge · ONNX specification
AI systems
Quantisation
INT8, INT4, FP8 and block-scaled formats that cut memory and power with measured accuracy loss.
We design and integrate.
OCP Microscaling Formats (MX) Specification v1.0 · IEEE 754-2019 (floating-point arithmetic)
AI systems
Distillation, pruning and sparsity
Training smaller student models and removing redundant weights so models fit edge hardware.
We design and integrate.
ISO/IEC 23053 (framework for AI systems using ML)
AI systems
Compilers and runtimes
Graph compilers and portable model formats that target many accelerator types from one model.
We design and integrate.
ONNX specification · MLIR (LLVM project)
AI systems
Serving: batching, caching and scheduling
Request batching, key-value caching and admission control that raise throughput on shared accelerators.
We design, integrate and operate.
MLPerf Inference
AI systems
Parameter-efficient fine-tuning
Adapting large models to a site, language or sensor with small trainable adapters.
We design and integrate.
ISO/IEC 23053
AI systems
Energy-aware inference
Measuring and reducing energy per inference as a design target alongside accuracy and latency.
We design and operate.
MLPerf Power · ISO/IEC 30134 (data centre KPIs)
Frontier · AI systems
Split and collaborative inference
Dividing a model between device, edge and cloud to balance latency, bandwidth and privacy.
We design and integrate.
3GPP TR 22.874 (AI/ML model transfer in 5GS) · 3GPP TR 23.700-80
02Architecture
Where it sits, and what it speaks.
Into the product draws on 4 domains and 32 layers of the spectrum; the highlighted layers carry this area. It speaks 3 of the protocol’s 8 core functions.
Edge inference, model serving and optimisation
10
Foundation models for Earth observation and geospatial
9
Time-series and sensor AI
8
Multimodal fusion and agentic operations
7
Federated and privacy-preserving learning
7
Physical AI and robotics (within red lines)
7
MLOps
7
Evaluation, model risk and AI governance
10
O-RAN architecture and open interfaces
7
Fronthaul (7.2x split) and transport
8
RAN intelligent controllers, xApps and rApps
8
SMO, O-Cloud and accelerated L1
8
AI-RAN: shared accelerated infrastructure
7
AI/ML in the air interface and network management
8
Energy saving and efficiency
6
Testing, integration and security
8
01JoinIdentity, attestation quote, signed policy.
02AdmitAdmitted for a stated scope, nothing wider.
03DescribeWhat, where, how and by whom, with calibration.
04TimeOne clock across the grid, with its error bound.
05ObserveEvery value with quality and provenance.
06QualifyTested against the specification, state on record.
07FederateSigned exchange across grids and trust domains.
08OrchestrateSoftware and configuration by signed plan.
03On the grid
Every site joins one observing grid.
The open networks on the grid today, over real relief and the geodesic cells we count them into, filtered to the space domain this area serves. A site built here enrols into the same grid.
04In the field
Where this runs, sector by sector.
6 deployment scenarios from the sector research that rely on into the product.
TransportNon-federal automated weather observing for small airports
Aviation-grade stations under the civil-aviation authority’s advisory circular; pricing not researched.Defence-adjacent (civil dual-use)Research and development route through the defence science programme
A science-and-technology programme on all-domain situational awareness.Government and public safetyNational GBON surface and upper-air network
WMO-grade stations reporting through WIS 2.0, funded through the systematic-observations financing pathway.Government and public safetyAlerting integration with the Common Alerting Protocol and cell broadcast
Software-heavy: sensor triggers feed an alert origination tool; the alert authority keeps every decision.Government and public safetyDeployable communications and sensing for incident command
A cell on wheels with satellite backhaul, mesh telemetry and environment nodes at the incident base.IndustrialUtility private LTE for the grid edge
Distribution automation, distributed-energy control and wildfire-risk weather on a dedicated band.05Open and enterprise
Open by default. Enterprise by contract.
The open counterpart lives on the main site for anyone to read and run. The enterprise service is designed, deployed and operated with you, under contract.
Open, on the main site
- Benchmark methods
- Pilot reports, published either way
Enterprise, on this site
- Research pilots
- Prototype engineering
06The rest of the line
Frontier labs, area by area.
Frontier labsPilots
New technology, a real site, a written plan.Frontier labsBenchmarks
Open data, one method, comparable results.Frontier labsAI-RAN and 6G
Sensing and communication on shared compute.Frontier labsSpace and NTN
Direct-to-device and new payloads, tested.Frontier labsPhysical AI
Drones and robots, inside published red lines.Bring us the site.
Survey, design, deployment and operation for into the product. The form names who receives your details.
Talk to uscontact@observatory.global